--- dataset_info: - config_name: filtered features: - name: text dtype: string - name: id dtype: string - name: dump dtype: string - name: url dtype: string - name: date dtype: string - name: file_path dtype: string - name: language dtype: string - name: language_score dtype: float64 - name: language_script dtype: string - name: minhash_cluster_size dtype: int64 - name: top_langs dtype: string splits: - name: train num_bytes: 14529270539 num_examples: 1619895 - name: test num_bytes: 143832829 num_examples: 16096 download_size: 4660212792 dataset_size: 14673103368 - config_name: removed features: - name: text dtype: string - name: id dtype: string - name: dump dtype: string - name: url dtype: string - name: date dtype: string - name: file_path dtype: string - name: language dtype: string - name: language_score dtype: float64 - name: language_script dtype: string - name: minhash_cluster_size dtype: int64 - name: filter_reason dtype: string - name: top_langs dtype: string splits: - name: train num_bytes: 5843948257 num_examples: 1033074 download_size: 1804153242 dataset_size: 5843948257 configs: - config_name: filtered data_files: - split: train path: filtered/train-* - split: test path: filtered/test-* - config_name: removed data_files: - split: train path: removed/train-* task_categories: - text-generation - fill-mask task_ids: - language-modeling - masked-language-modeling language: - my pretty_name: Myanmar Fineweb2 Dataset --- *Please visit to the [GitHub repository](https://github.com/chuuhtetnaing/myanmar-language-dataset-collection) for other Myanmar Langauge datasets.* # Myanmar Fineweb2 Dataset A preprocessed subset of the [Fineweb2 dataset](https://huggingface.co/datasets/HuggingFaceFW/fineweb-2) containing only Myanmar language text, with consistent **Unicode** encoding. ## Dataset Description This dataset is derived from the [Fineweb2](https://huggingface.co/datasets/HuggingFaceFW/fineweb-2) created by [HuggingFaceFW](https://huggingface.co/HuggingFaceFW). It contains only the Myanmar language portion of the original Fineweb2 dataset, with additional preprocessing to standardize text encoding. ## Filtered and Removed Subsets This dataset provides two configurations: - **filtered**: Contains Myanmar text that passed the original filtering criteria - **removed**: Contains Myanmar text that was filtered out by the original filtering pipeline These two subsets together represent the complete Myanmar language data from Fineweb2 after global deduplication. This structure allows researchers to easily apply their own filtering criteria or work with both the filtered and unfiltered data according to their needs. > Quote from original dataset: > While we tried our best to not overfilter, we know that our filtering isn't perfect, and wanted to allow the community to **easily re-filter the data with their own filtering criteria**. We have therefore also uploaded the data that was **removed** by our filtering pipeline for each language (it is suffixed by `_removed`). The _filtered + the removed subsets_ of each language represent the entire data for a given language following global deduplication, which means that you do not have to re-deduplicate it yourself. You can find and adapt our filtering [code here](https://github.com/huggingface/fineweb-2/blob/main/fineweb-2-pipeline.py). ### Preprocessing The main preprocessing step applied to this dataset was: - **Zawgyi to Unicode conversion**: Myanmar text can be encoded in two different ways - Zawgyi and Unicode. We detected Zawgyi-encoded text and converted it to Unicode for consistency, ensuring all text in the dataset uses the same encoding standard. The conversion was performed using [Myanmar Tools](https://github.com/google/myanmar-tools) for detection and ICU for transliteration: ```python from myanmartools import ZawgyiDetector from icu import Transliterator # Initialize the detector and converter detector = ZawgyiDetector() converter = Transliterator.createInstance('Zawgyi-my') # Example conversion function def zawgyi_to_unicode(text): score = detector.get_zawgyi_probability(text) if score > 0.5: # If likely Zawgyi return converter.transliterate(text) return text # Already Unicode ``` ### Dataset Structure The dataset keep the same fields as the original fineweb-2 dataset. ## Usage ### Load Filtered Subset You can load this dataset's filtered subset using the Hugging Face datasets library: ```python from datasets import load_dataset dataset = load_dataset("chuuhtetnaing/myanmar-c4-dataset") ``` ### Load Whole Dataset (including removed) You can also load this whole dataset using the Hugging Face datasets library: ```python from datasets import concatenate_datasets filtered_ds = load_dataset("chuuhtetnaing/myanmar-fineweb-2-dataset", name='filtered') removed_ds = load_dataset("chuuhtetnaing/myanmar-fineweb-2-dataset", name='removed') ds = DatasetDict({ "train": concatenate_datasets([filtered_ds["train"], removed_ds["train"]]), "test": filtered_ds["test"] }) ``` ## Dataset Creation This dataset was created by: 1. Extracting the Myanmar language split (both filtered and removed) from the original Fineweb2 dataset 2. Detecting Zawgyi-encoded text using Google's Myanmar Tools probabilistic detector on each line 3. Converting Zawgyi text to Unicode using ICU's transliteration converter on each line ## Dependencies The preprocessing of this dataset relied on: - [Myanmar Tools](https://github.com/google/myanmar-tools) - For Zawgyi detection - [PyICU](https://pypi.org/project/PyICU/) - For Zawgyi to Unicode conversion - [Documentation](https://github.com/google/myanmar-tools/blob/master/clients/python/README.rst) - Documentation from Myanmar Tools ## License This dataset follows the same license as the original Fineweb2 dataset. Please refer to the [original dataset page](https://huggingface.co/datasets/HuggingFaceFW/fineweb-2) for licensing information. Myanmar Tools is released under the [Apache License 2.0](https://github.com/google/myanmar-tools/blob/master/LICENSE).